# Limits and retries

Keep batches, polling, and account capacity within the service contract.

## Account and request limits

| Operation                                  | Current bound                               |
| ------------------------------------------ | ------------------------------------------- |
| Queued or processing downloads per account | At most 10, or a lower configured ceiling   |
| New ordinary download requests             | 20 per minute per account                   |
| Job status reads                           | 300 per minute per account                  |
| HTTP batch or playlist download selection  | 1 to 10 items                               |
| Saved playlists                            | 5 per account                               |
| Saved links                                | 50 per playlist                             |
| Items in an Instagram post result          | Up to 50                                    |
| Private signed URL lifetime                | Normally 900 seconds; operator-configurable |

Keys on the same account do not multiply these allowances. The authenticated artifact/download allowance defaults to 5 GiB, subject to worker and deployment ceilings. A post's byte limit is shared across its collection. These bounds are not a promise of sustained bandwidth or weighted partner scheduling.

## Polling

Poll jobs about once every two seconds and stop on a terminal status. Use bounded waiting in a UI or service. Use SSE when it fits your client, and fall back to status lookup after a reconnect.

## Retry safely

For `429` or a temporary `5xx`, wait and increase the delay between retries. Honor `Retry-After` when supplied. Use jitter when several workers could retry together.

For a timed-out individual submission, keep its `Idempotency-Key`. Reusing that key for a different body returns `409`. Do not blindly resubmit batches, because HTTP batch creation does not have the same idempotency contract.

An expired key is not repaired by retrying it. A permanent source failure, invalid request, or ownership failure also needs a change rather than an immediate retry loop.